This implementation is based on InterfaceGAN (CVPR 2020). In lieu of utilizing StyleGANv1 within InterfaceGAN, I have customized the codebase to align with StyleGANv2. This adaptation ensures compatibility and leverages the enhanced features offered by StyleGANv2.

Download the model checkpoints from here. Place them to ./checkpoints.

Model checkpoints include e4e_ffhq_encode.pt, shape_predictor_68_face_landmarks.dat, stylegan2-ffhq-config-f.pt.

I have tested on:

• torch (PyTorch 1.13.0)
• torchvision (0.14.0)
• CUDA 11.6

• Train an attribute classifier based on your specified attributes.
• Place the pretrained classifier to ./checkpoints/classifiers and name it [ATTRIBUTUE].pth.

• Prepare the data for attribute vector solving.

python preparing.py --attr=[ATTRIBUTE] --n_samples=20000

python solve.py --attr=[ATTRIBUTE] --code=w

• The solved attribute vectors will be saved to ./checkpoints/attribute_vectors.

Adjust the placeholders such as [ATTRIBUTE] according to your specific attribute names.

Single attribute manipulation only.

python manipulation.py --attr=[ATTRIBUTE]

Results will be saved to ./outputs.

Provide a facial image for semantic editing. Make sure checkpoints are well-prepared.

python inference.py --input=[IMAGE_PATH] --attr=[ATTRIBUTE] --alpha=3 --conditions=[ATTRIBURES(optinal)]

Results will be saved to ./outputs.

The StyleGANv2 is borrowed from this pytorch implementation by @rosinality. The implementation of e4e projection is also heavily from encoder4editing.